Practice 04 · Artificial intelligence & automationCustom AI assistant: a tool that works with your documents and speaks your vocabulary
General-purpose assistants write well, but they know nothing of your standard contracts, your product nomenclature or the unwritten rules of your trade. A custom AI assistant closes that gap: connected to your own sources and configured for identified tasks, it shows where each part of its answer comes from. It does not replace the expert; it prepares the ground for them.
A 30-minute conversation about your uses, free
Who it is for, and whenWhen a custom AI assistant is worth building
The need appears when skilled staff spend hours reading large volumes of text, pulling together scattered information or producing documents that always follow the same logic.
- An engineering office or consultancy must analyse long specifications or regulations and pull out the points that concern your offer.
- A sales department writes proposals that reuse existing material (references, technical descriptions, terms) adapted to each prospect.
- A quality or legal team compares incoming documents against internal standards and flags deviations before a decision.
- Management prepares its committees from reports scattered across several tools.
Under the hoodThe path of a question, from your documents to a sourced answer
The language model never sees all of your documents: it only receives the few passages retrieved for the question asked, and only those the person is allowed to open. When no passage fits, the assistant says so instead of filling the gap itself.
Upstream, whenever the sources change
Your documents
Contracts, procedures, product sheets, project histories: the admitted sources, defined with you.
- Contracts
- Procedures
- Product sheets
- Projects
Splitting and indexing
Each document becomes passages, indexed by meaning and carrying their date, origin and access rights.
For each question
Question
“What late-delivery penalties does our supplier framework agreement set?”
Asked by: Purchasing
Search filtered by access rights
The passages closest to the question, among those the person can open.
- Supplier framework agreement, cl. 9.2retained
- Amendment no. 2, cl. 3retained
- Finance department memono access
- General terms of salenot relevant
Language model
Instruction: answer only from the retained passages, cite them, flag what is missing.
Sourced answer
The framework agreement sets a penalty of 0.5% of the order value per day of delay, capped at 10%.1 Amendment no. 2 excludes delays caused by the carrier.2
- 1Framework agreement, cl. 9.2
- 2Amendment no. 2, cl. 3
No relevant passage
“I cannot find anything on this point in the documents you have access to.”
Same reflex when two sources contradict each other: they are shown side by side.
The scopeWhat Kopus designs and delivers
The assistant relies on a search mechanism over your documents (contracts, procedures, product sheets, project histories) that feeds the language model with the relevant passages at the moment of the question. This architecture, known as retrieval-augmented generation, allows it to answer from your content and cite the documents it used. We define with you which sources are admitted and which are excluded.
We then configure specific tasks as working templates: produce an analysis note following your usual outline, fill in a comparison grid, prepare a first draft of a proposal, extract the obligations from a contract. Each task carries a caution instruction: flag passages with no support in the sources rather than fill the gaps.
Access rules mirror those of your documents: a user sees in the answers only what they could open themselves. The model provider and hosting location depend on how sensitive your data is; European hosting or your own infrastructure are among the options.
What you get
- An assistant that speaks your business language and follows your rules.
- Sourced, verifiable, up-to-date answers.
- Real time savings for your teams and your customers.
The methodHow we build a custom AI assistant
Use case framing
We select two or three high-value tasks, described by the people who perform them, with examples of the expected output.
Source inventory
Which documents, in which tools, under which access rules, and which portion is current enough to serve as a reference.
Prototype evaluated on real cases
The assistant processes past files whose correct answer you know, compared against your experts’ results.
Tuning of instructions and sources
Rewording of working templates, adding or removing documents, calibrating the uncertainty signals.
Rollout to the chosen user group, with a channel for reporting disappointing answers.
Our positionOur position: an assistant must show its sources or admit its doubts
A fluent, well-written answer inspires confidence, including when it is wrong. That is the main danger of an assistant working on subjects where mistakes are expensive. Every assistant we design therefore points back to the exact passages behind its answer, and says when information is missing or when two documents contradict each other. An expert who can verify in one click works faster; an expert who has to re-check everything has gained nothing.
Your questionsFrequently asked questions about this service
When should you build a custom AI assistant rather than use an off-the-shelf tool?
When your documents, vocabulary and rules go beyond what a general assistant can know, or when your confidentiality constraints rule out sending your data to a consumer service.
Does the assistant keep learning from our documents?
It does not modify the language model; it consults your documents at each question. An update to the sources is therefore reflected immediately.
How does this differ from an AI assistant for employees?
The custom assistant is a production tool for precise analysis and drafting tasks, often reserved for a few roles. The employee assistant answers the everyday questions of the whole company on procedures, HR or IT.
Can the assistant work with confidential data?
Yes, provided the right architecture is chosen: suitable hosting, partitioned access, no training on your data. We define this with you from the framing stage.
Do we need to rewrite our documents for the assistant?
Generally not. We work from your existing documents and flag those that would benefit from clarification, which also helps your teams.
How long does it take to build a custom AI assistant?
Three to eight weeks depending on the volume of knowledge to prepare and the integrations, including an evaluation phase with your experts before go-live.
Go furtherPages related to this service
Our workProjects built on this expertise
All our work
Let’s bring your projects to lifeYour next project starts here.
Let’s define the tasks where an assistant connected to your data would save you the most.
A 30-minute conversation about your uses, free